Skip to content
Snippets Groups Projects
Commit 4b742e28 authored by Mirzam98's avatar Mirzam98
Browse files

[fpgadataflow/convinpgen]: added support for non-square images

parent e68d7a3a
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
+ 69
− 49
View file @ 4b742e28
...@@ -61,12 +61,13 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -61,12 +61,13 @@ class ConvolutionInputGenerator(HLSCustomOp):
def get_nodeattr_types(self): def get_nodeattr_types(self):
my_attrs = { my_attrs = {
"ConvKernelDim": ("i", True, 0), "ConvKernelDim": ("ints", True, []), # [H, W] = [Y, X]
"IFMChannels": ("i", True, 0), "IFMChannels": ("i", True, 0),
"IFMDim": ("i", True, 0), "IFMDim": ("ints", True, []), # [H, W] = [Y, X]
"OFMDim": ("i", True, 0), "OFMDim": ("ints", True, []), # [H, W] = [Y, X]
"SIMD": ("i", True, 0), "SIMD": ("i", True, 0),
"Stride": ("i", True, 0), "Stride": ("ints", True, [1, 1]), # [H, W] = [Y, X]
"Dilation": ("ints", True, [1, 1]), # [H, W] = [Y, X]
# FINN DataTypes for inputs, weights, outputs # FINN DataTypes for inputs, weights, outputs
"inputDataType": ("s", True, ""), "inputDataType": ("s", True, ""),
"outputDataType": ("s", True, ""), "outputDataType": ("s", True, ""),
...@@ -87,43 +88,45 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -87,43 +88,45 @@ class ConvolutionInputGenerator(HLSCustomOp):
return my_attrs return my_attrs
def get_normal_input_shape(self): def get_normal_input_shape(self):
ifm_dim_h, ifm_dim_w = self.get_nodeattr("IFMDim")
ifm_dim = self.get_nodeattr("IFMDim")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
ishape = (1, ifm_dim_h, ifm_dim_w, ifm_ch)
ishape = (1, ifm_dim, ifm_dim, ifm_ch)
return ishape return ishape
def get_folded_input_shape(self): def get_folded_input_shape(self):
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim_h, ifm_dim_w = self.get_nodeattr("IFMDim")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
assert ifm_ch % simd == 0, "SIMD must divide IFMChannels" assert ifm_ch % simd == 0, "SIMD must divide IFMChannels"
wf = int(ifm_ch / simd) wf = int(ifm_ch / simd)
folded_ishape = (1, ifm_dim, ifm_dim, wf, simd) folded_ishape = (1, ifm_dim_h, ifm_dim_w, wf, simd)
return folded_ishape return folded_ishape
def get_normal_output_shape(self): def get_normal_output_shape(self):
k = self.get_nodeattr("ConvKernelDim") k_h, k_w = self.get_nodeattr("ConvKernelDim")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim_h, ifm_dim_w = self.get_nodeattr("IFMDim")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
stride = self.get_nodeattr("Stride") stride_h, stride_w = self.get_nodeattr("Stride")
dilation_h, dilation_w = self.get_nodeattr("Dilation")
pad = 0 pad = 0
ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad) ofm_dim_h = compute_conv_output_dim(ifm_dim_h, k_h, stride_h, pad, dilation_h)
oshape = (1, ofm_dim, ofm_dim, k * k * ifm_ch) ofm_dim_w = compute_conv_output_dim(ifm_dim_w, k_w, stride_w, pad, dilation_w)
oshape = (1, ofm_dim_h, ofm_dim_w, k_h * k_w * ifm_ch)
return oshape return oshape
def get_folded_output_shape(self): def get_folded_output_shape(self):
k = self.get_nodeattr("ConvKernelDim") k_h, k_w = self.get_nodeattr("ConvKernelDim")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim_h, ifm_dim_w = self.get_nodeattr("IFMDim")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
stride = self.get_nodeattr("Stride") stride_h, stride_w = self.get_nodeattr("Stride")
dilation_h, dilation_w = self.get_nodeattr("Dilation")
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
pad = 0 pad = 0
ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad) ofm_dim_h = compute_conv_output_dim(ifm_dim_h, k_h, stride_h, pad, dilation_h)
ofm_dim_w = compute_conv_output_dim(ifm_dim_w, k_w, stride_w, pad, dilation_w)
assert ifm_ch % simd == 0, "SIMD must divide IFMChannels" assert ifm_ch % simd == 0, "SIMD must divide IFMChannels"
wf = int((k * k * ifm_ch) // simd) wf = int((k_h * k_w * ifm_ch) // simd)
folded_oshape = (1, ofm_dim, ofm_dim, wf, simd) folded_oshape = (1, ofm_dim_h, ofm_dim_w, wf, simd)
return folded_oshape return folded_oshape
def make_shape_compatible_op(self, model): def make_shape_compatible_op(self, model):
...@@ -186,26 +189,31 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -186,26 +189,31 @@ class ConvolutionInputGenerator(HLSCustomOp):
def get_exp_cycles(self): def get_exp_cycles(self):
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
k = self.get_nodeattr("ConvKernelDim") k_h, k_w = self.get_nodeattr("ConvKernelDim")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim_h, ifm_dim_w = self.get_nodeattr("IFMDim")
ofm_dim = self.get_nodeattr("OFMDim") ofm_dim_h, ofm_dim_w = self.get_nodeattr("OFMDim")
stride = self.get_nodeattr("Stride") stride_h, stride_w = self.get_nodeattr("Stride")
dilation_h, dilation_w = self.get_nodeattr("Dilation")
# since mmv != 1 is not supported yet, we set mmv for now to 1 # since mmv != 1 is not supported yet, we set mmv for now to 1
mmv = 1 mmv = 1
# see https://github.com/Xilinx/finn-hlslib/blob/master/slidingwindow.h # see https://github.com/Xilinx/finn-hlslib/blob/master/slidingwindow.h
cycles_write_block = (ofm_dim * k * k * (ifm_ch / simd)) / mmv cycles_write_block = (ofm_dim_w * k_w * k_h * (ifm_ch / simd)) / mmv
cycles_read_block = stride * ifm_dim * (ifm_ch / simd) cycles_read_block = stride_w * ifm_dim_w * (ifm_ch / simd)
max_cycles = max(cycles_write_block, cycles_read_block) max_cycles = max(cycles_write_block, cycles_read_block)
exp_cycles = ifm_dim * k * (ifm_ch / simd) + ofm_dim * max_cycles exp_cycles = (
ifm_dim_w * k_h * dilation_h * (ifm_ch / simd) + ofm_dim_h * max_cycles
)
return int(exp_cycles) return int(exp_cycles)
def bram_estimation(self): def bram_estimation(self):
# NOTE: only tested with a square convolution
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim = self.get_nodeattr("IFMDim")[0]
k = self.get_nodeattr("ConvKernelDim") k = self.get_nodeattr("ConvKernelDim")[0]
stride = self.get_nodeattr("Stride") stride = self.get_nodeattr("Stride")[0]
ram_style = self.get_nodeattr("ram_style") ram_style = self.get_nodeattr("ram_style")
if ram_style == "block" or ram_style == "auto": if ram_style == "block" or ram_style == "auto":
ram_depth = ifm_dim * ifm_ch / simd ram_depth = ifm_dim * ifm_ch / simd
...@@ -232,11 +240,12 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -232,11 +240,12 @@ class ConvolutionInputGenerator(HLSCustomOp):
return 0 return 0
def lut_estimation(self): def lut_estimation(self):
# NOTE: only tested with a square convolution
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim = self.get_nodeattr("IFMDim")[0]
k = self.get_nodeattr("ConvKernelDim") k = self.get_nodeattr("ConvKernelDim")[0]
stride = self.get_nodeattr("Stride") stride = self.get_nodeattr("Stride")[0]
ram_style = self.get_nodeattr("ram_style") ram_style = self.get_nodeattr("ram_style")
if ram_style == "distributed": if ram_style == "distributed":
ram_luts = int( ram_luts = int(
...@@ -252,11 +261,12 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -252,11 +261,12 @@ class ConvolutionInputGenerator(HLSCustomOp):
return 300 + ram_luts return 300 + ram_luts
def uram_estimation(self): def uram_estimation(self):
# NOTE: only tested with a square convolution
simd = self.get_nodeattr("SIMD") simd = self.get_nodeattr("SIMD")
ifm_ch = self.get_nodeattr("IFMChannels") ifm_ch = self.get_nodeattr("IFMChannels")
ifm_dim = self.get_nodeattr("IFMDim") ifm_dim = self.get_nodeattr("IFMDim")[0]
k = self.get_nodeattr("ConvKernelDim") k = self.get_nodeattr("ConvKernelDim")[0]
stride = self.get_nodeattr("Stride") stride = self.get_nodeattr("Stride")[0]
ram_style = self.get_nodeattr("ram_style") ram_style = self.get_nodeattr("ram_style")
if ram_style == "ultra": if ram_style == "ultra":
return int( return int(
...@@ -295,7 +305,7 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -295,7 +305,7 @@ class ConvolutionInputGenerator(HLSCustomOp):
assert ( assert (
inp.shape == exp_ishape inp.shape == exp_ishape
), """Input shape doesn't ), """Input shape doesn't
match expected shape (1, ifm_dim, ifm_dim, ifm_ch).""" match expected shape (1, ifm_dim_h, ifm_dim_w, ifm_ch)."""
if self.get_input_datatype() == DataType.BIPOLAR: if self.get_input_datatype() == DataType.BIPOLAR:
# store bipolar activations as binary # store bipolar activations as binary
inp = (inp + 1) / 2 inp = (inp + 1) / 2
...@@ -354,25 +364,33 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -354,25 +364,33 @@ class ConvolutionInputGenerator(HLSCustomOp):
assert ( assert (
context[node.output[0]].shape == exp_oshape context[node.output[0]].shape == exp_oshape
), """Output ), """Output
shape doesn't match expected shape (1, ofm_dim, ofm_dim, k*k*ifm_ch).""" shape doesn't match expected shape (1, ofm_dim_h, ofm_dim_w, k_h*k_w*ifm_ch)."""
def global_includes(self): def global_includes(self):
self.code_gen_dict["$GLOBALS$"] = ['#include "slidingwindow.h"'] self.code_gen_dict["$GLOBALS$"] = ['#include "slidingwindow.h"']
def defines(self, var): def defines(self, var):
numReps = 1 numReps = 1
ifm_dim = self.get_nodeattr("IFMDim")[0]
ifm_ch = self.get_nodeattr("IFMChannels")
ofm_dim = self.get_nodeattr("OFMDim")[0]
k = self.get_nodeattr("ConvKernelDim")[0]
stride = self.get_nodeattr("Stride")[0]
simd = self.get_nodeattr("SIMD")
ifm_precision = self.get_input_datatype().bitwidth()
self.code_gen_dict["$DEFINES$"] = [ self.code_gen_dict["$DEFINES$"] = [
"""#define ConvKernelDim1 {}\n #define IFMChannels1 {}\n """#define ConvKernelDim1 {}\n #define IFMChannels1 {}\n
#define Input_precision1 {}\n #define IFMDim1 {}\n #define Input_precision1 {}\n #define IFMDim1 {}\n
#define OFMDim1 {}\n #define SIMD1 {}\n #define OFMDim1 {}\n #define SIMD1 {}\n
#define Stride1 {}\n #define numReps {}""".format( #define Stride1 {}\n #define numReps {}""".format(
self.get_nodeattr("ConvKernelDim"), k,
self.get_nodeattr("IFMChannels"), ifm_ch,
self.get_input_datatype().bitwidth(), ifm_precision,
self.get_nodeattr("IFMDim"), ifm_dim,
self.get_nodeattr("OFMDim"), ofm_dim,
self.get_nodeattr("SIMD"), simd,
self.get_nodeattr("Stride"), stride,
numReps, numReps,
) )
] ]
...@@ -415,9 +433,11 @@ class ConvolutionInputGenerator(HLSCustomOp): ...@@ -415,9 +433,11 @@ class ConvolutionInputGenerator(HLSCustomOp):
} }
hls_ram_style = map_to_hls_ram_style[ram_style] hls_ram_style = map_to_hls_ram_style[ram_style]
hls_call = node.op_type hls_call = node.op_type
# check if non optimized ConvolutionInputGenerator is needed
k = self.get_nodeattr("ConvKernelDim") # check which ConvolutionInputGenerator is needed
stride = self.get_nodeattr("Stride") k = self.get_nodeattr("ConvKernelDim")[0]
stride = self.get_nodeattr("Stride")[0]
if k % stride != 0: if k % stride != 0:
hls_call += "_kernel_stride" hls_call += "_kernel_stride"
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment